huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-21 20:22:49 +00:00
Code Issues Packages Projects Releases Wiki Activity

Upsample grad (#1420)

Browse Source 
* encode size of upsample in enum

* working convolution method for limited 2d kernels

* add test for sf 3 interpolation

* add higher dimensional tests, fix to work with multichannel input

* Remove commented out line.

---------

Co-authored-by: Laurent <laurent.mazare@gmail.com>
This commit is contained in:
KGrewal1
2023-12-10 07:43:24 +00:00
committed by GitHub
parent 9bd94c1ffa
commit 18eb87f25f
4 changed files with 192 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

160
candle-core/tests/grad_tests.rs
View File

@ -270,6 +270,166 @@ fn unary_grad(device: &Device) -> Result<()> {
[0.7358, 2.0000, 0.2707, 1.0000]
);
// manually checked: see comments
let x = Var::new(&[[[[1f32, 2., 3.], [4., 5., 6.], [7., 8., 9.]]]], device)?;
let y = x.interpolate2d(6, 6)?.reshape(36)?;
#[rustfmt::skip]
let z = Tensor::new(
&[
1_f32, 02., 03., 04., 05., 06.,
07., 08., 09., 10., 11., 12.,
13., 14., 15., 16., 17., 18.,
19., 20., 21., 22., 23., 24.,
25., 26., 27., 28., 29., 30.,
31., 32., 33., 34., 35., 36.,
],
device,
)?;
// gradient should be
// row 1
// 1+2+7+8 = 18
// 3+4+9+10 = 26
// 5+6+11+12 = 34
// row 2
// 13+14+19+20 = 66
// 15+16+21+22 = 74
// 17+18+23+24 = 82
// row 3
// 25+26+31+32 = 114
// 27+28+33+34 = 122
// 29+30+35+36 = 130
let loss = y.unsqueeze(1)?.transpose(0, 1)?.matmul(&z.unsqueeze(1)?)?;
let grads = loss.backward()?;
let grad_x = grads.get(&x).context("no grad for x")?;
assert_eq!(
test_utils::to_vec2_round(&grad_x.flatten(0, 2)?, 4)?,
[[18_f32, 26., 34.], [66., 74., 82.], [114., 122., 130.]]
);
// manually checked: see comments
let x = Var::new(&[[[[1f32, 2.], [4., 5.]]]], device)?;
let y = x.interpolate2d(6, 6)?.reshape(36)?;
#[rustfmt::skip]
let z = Tensor::new(
&[
1_f32, 02., 03., 04., 05., 06.,
07., 08., 09., 10., 11., 12.,
13., 14., 15., 16., 17., 18.,
19., 20., 21., 22., 23., 24.,
25., 26., 27., 28., 29., 30.,
31., 32., 33., 34., 35., 36.,
],
device,
)?;
// gradient should be
// row 1
// 1+2+3+7+8+9+13+14+15 = 72
// 4+5+6+10+11+12+16+17+18 = 99
// row 2
// 19+20+21+25+26+27+31+32+33 = 234
// 22+23+24+28+29+30+34+35+36 = 243
let loss = y.unsqueeze(1)?.transpose(0, 1)?.matmul(&z.unsqueeze(1)?)?;
let grads = loss.backward()?;
let grad_x = grads.get(&x).context("no grad for x")?;
assert_eq!(
test_utils::to_vec2_round(&grad_x.flatten(0, 2)?, 4)?,
[[72_f32, 99.], [234., 261.]]
);
// manually checked: see comments
let x = Var::new(&[[[[1f32, 2.], [4., 5.]], [[6f32, 7.], [8., 9.]]]], device)?;
let y = x.interpolate2d(4, 4)?.reshape(32)?;
#[rustfmt::skip]
let z = Tensor::new(
&[
1_f32, 02., 03., 04.,
05., 06., 07., 08.,
09., 10., 11., 12.,
13., 14., 15., 16.,
17., 18., 19., 20.,
21., 22., 23., 24.,
25., 26., 27., 28.,
29., 30., 31., 32.
],
device,
)?;
// gradient should be
// m1r1
// 1+2+5+6=14
// 3+4+7+8=22
// m1r2
// 9+10+13+14=46
// 11+12+15+16=54
// m2r1
// 17+18+21+22=78
// 19+20+23+24=86
// m2r2
// 25+26+29+30=110
// 27+28+31+32=118
let loss = y.unsqueeze(1)?.transpose(0, 1)?.matmul(&z.unsqueeze(1)?)?;
let grads = loss.backward()?;
let grad_x = grads.get(&x).context("no grad for x")?;
assert_eq!(
test_utils::to_vec3_round(&grad_x.flatten(0, 1)?, 4)?,
[[[14_f32, 22.], [46., 54.]], [[78., 86.], [110., 118.]]]
);
// manually checked: see comments
let x = Var::new(
&[[[[1f32, 2.], [4., 5.]]], [[[6f32, 7.], [8., 9.]]]],
device,
)?;
let y = x.interpolate2d(4, 4)?.reshape(32)?;
#[rustfmt::skip]
let z = Tensor::new(
&[
1_f32, 02., 03., 04.,
05., 06., 07., 08.,
09., 10., 11., 12.,
13., 14., 15., 16.,
17., 18., 19., 20.,
21., 22., 23., 24.,
25., 26., 27., 28.,
29., 30., 31., 32.
],
device,
)?;
// gradient should be
// m1r1
// 1+2+5+6=14
// 3+4+7+8=22
// m1r2
// 9+10+13+14=46
// 11+12+15+16=54
// m2r1
// 17+18+21+22=78
// 19+20+23+24=86
// m2r2
// 25+26+29+30=110
// 27+28+31+32=118
let loss = y.unsqueeze(1)?.transpose(0, 1)?.matmul(&z.unsqueeze(1)?)?;
let grads = loss.backward()?;
let grad_x = grads.get(&x).context("no grad for x")?;
assert_eq!(
test_utils::to_vec3_round(&grad_x.flatten(0, 1)?, 4)?,
[[[14_f32, 22.], [46., 54.]], [[78., 86.], [110., 118.]]]
);
Ok(())
}